1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
|
#include <torch/csrc/jit/runtime/static/impl.h>
#include <ATen/core/interned_strings.h>
#include <ATen/core/op_registration/op_registration.h>
#include <torch/csrc/jit/passes/canonicalize.h>
#include <torch/csrc/jit/passes/freeze_module.h>
#include <torch/csrc/jit/passes/remove_mutation.h>
#include <torch/csrc/jit/passes/subgraph_rewrite.h>
#include <torch/csrc/jit/runtime/static/ops.h>
#include <torch/csrc/jit/runtime/vararg_functions.h>
namespace torch {
namespace jit {
using c10::DispatchKey;
using c10::RegisterOperators;
std::shared_ptr<torch::jit::Graph> PrepareForStaticRuntime(
std::shared_ptr<torch::jit::Graph> g) {
Inline(*g);
ConstantPropagation(g);
Canonicalize(g);
ConstantPropagation(g);
RemoveTensorMutation(g);
ConstantPropagation(g);
for (auto n : g->nodes()) {
if (n->kind() == c10::Symbol::fromQualString("prim::GetAttr")) {
throw std::runtime_error("Cannot accelerate unfrozen graphs");
}
bool supported = false;
#define X(_) \
if (n->kind() == c10::Symbol::fromQualString(#_)) { \
supported = true; \
}
SUPPORTED_OPS(X)
#undef X
if (!supported) {
throw std::runtime_error(
std::string("Unsupported operation: ") + n->kind().toQualString());
}
}
// remove unused input 0 from graph
if (g->inputs().at(0)->type()->is_module()) {
if (!g->inputs().at(0)->hasUses()) {
g->eraseInput(0);
}
}
return g;
}
std::shared_ptr<torch::jit::Graph> PrepareForStaticRuntime(
const torch::jit::Module& m) {
auto module = m.copy();
module.eval();
module = freeze_module(module);
auto g = module.get_method("forward").graph();
return PrepareForStaticRuntime(g);
}
StaticRuntime::StaticRuntime(std::shared_ptr<torch::jit::Graph> g) : graph_(g) {
// fill workspace_ with constants
for (Node* node : graph_->nodes()) {
if (node->kind() == prim::Constant) {
CHECK(node->output()->type()->kind() != FunctionType::Kind);
workspace_[node->output()] = toIValue(node->output()).value();
} else {
nodes_.emplace_back(node);
}
}
}
std::vector<at::Tensor> StaticRuntime::run(
const std::vector<at::Tensor>& inps) const {
// Container for inputs, outputs, and activations (excluding parameters)
TORCH_INTERNAL_ASSERT(graph_->inputs().size() == inps.size());
for (size_t i = 0; i < inps.size(); i++) {
workspace_[graph_->inputs()[i]] = inps[i];
}
for (const auto& n : nodes_) {
n.run(workspace_);
}
std::vector<at::Tensor> out;
for (Value* output : graph_->outputs()) {
const IValue& v = workspace_[output];
if (v.isTuple()) {
auto t = v.toTuple();
for (const auto& el : t->elements()) {
out.emplace_back(el.toTensor());
}
} else {
out.emplace_back(v.toTensor());
}
}
return out;
}
ProcessedNode::ProcessedNode(Node* node) : node_(node) {
if (node->kind() != prim::ListConstruct &&
node->kind() != prim::TupleConstruct) {
const Operator& op = node->getOperator();
CHECK(op.hasOperation());
op_ = op.getOperation(node);
}
if (canRunOutOfPlace(node)) {
fn_ = getOutOfPlaceOperation(node);
}
}
void ProcessedNode::run(StaticRuntime::ConstantMap& workspace) const {
if (!fn_) {
std::vector<IValue> stack;
const size_t size = node_->inputs().size();
stack.reserve(size);
for (size_t i = 0; i < size; i++) {
Value* v = node_->inputs()[i];
auto f = workspace.find(v);
TORCH_CHECK(
f != workspace.end(),
"Workspace does not contain Value ",
v->debugName());
stack.emplace_back(f->second);
}
if (op_) {
op_->operator()(&stack);
} else {
if (node_->kind() == prim::ListConstruct) {
listConstruct(
stack,
node_->output()->type()->expect<ListType>(),
node_->inputs().size());
} else if (node_->kind() == prim::TupleConstruct) {
bool named =
node_->output()->type()->expect<TupleType>()->name().has_value();
if (named) {
namedTupleConstruct(
stack,
node_->output()->type()->expect<TupleType>(),
node_->inputs().size());
} else {
tupleConstruct(stack, node_->inputs().size());
}
} else {
TORCH_CHECK(0, "Unhandled operation!", node_->kind().toQualString());
}
}
DCHECK_EQ(stack.size(), node_->outputs().size());
for (auto i = 0; i < node_->outputs().size(); i++) {
workspace[node_->outputs()[i]] = stack[i];
}
} else {
fn_->operator()(workspace);
}
}
} // namespace jit
} // namespace torch
|
